Activity Recognition in a Dense Sensor Network

A dense sensor network consisting of passive infrared motion detectors was developed and used to record human activity in hallways and rooms in a large campus building. Algorithms were developed that: (a) automatically determine the topology of the network from the sensor data, so that manual mapping is not required, (b) automatically learn patterns of sensor readings in local spatial and temporal neighborhoods, and (c) use the distribution of local events over larger spatial and temporal scales to automatically discover patterns of underlying global activities. The statistical distribution of the local events is analyzed using Probabilistic Latent Semantic Analysis (pLSA). Preliminary results show that the method can identify “typical” and “anomalous” activities.